AGGAT-based electromagnetic spectrum sensing open set target identification method

Abstract

The invention discloses an AGGAT-based electromagnetic spectrum sensing open set target recognition method, which comprises the following steps: firstly, constructing a knowledge graph representing formation information based on target multi-dimensional features extracted from an electromagnetic spectrum, and converting the knowledge graph into a representation form under a graph model through data coding and normalization; a global attention mechanism is introduced to construct an adaptive global graph attention network, comprehensive utilization of local information and global information in the graph is realized, feature extraction of targets in the formation graph is realized, and AGGAT is applied to feature extraction of the targets; designing a target type recognition loss function comprehensively utilizing supervised learning and comparative learning, and applying the target type recognition loss function to a training process of target type feature updating; and finally, constructing a feature clustering model of a density-based anti-noise clustering method, and constructing a feature comparison model based on Euclidean distance for outputting the type of the open set target. According to the method, the information source can be adaptively selected under different data distributions, so that the flexibility and characterization capability of image attention model feature extraction are improved.

Description

Technical Field

[0001] The present invention belongs to the field of cognitive reconnaissance of information and communication engineering, and specifically relates to an electromagnetic spectrum perception open set target recognition method based on AGGAT. Background Art

[0002] In electronic warfare, target recognition is a core element to improve situational awareness, enhance confrontation effectiveness and ensure mission success. Electromagnetic waves are an important carrier for information acquisition and transmission in electronic warfare. By deeply mining the electromagnetic spectrum information, the communication characteristics, radar signal patterns and movement laws of the target can be obtained, thereby providing auxiliary information for target recognition in various confrontation scenarios at sea, land and air. The traditional closed set target recognition algorithm assumes that all samples to be tested come from known categories. It extracts the electromagnetic information of the target and matches the features with the known target, which faces great limitations in practical applications. When facing new targets and unknown devices, the recognition model is required to have open set processing capabilities to ensure accurate identification of known targets and effective response to unknown targets. However, due to environmental interference and the inherent limitations of detection equipment such as radar and software-defined radio, the electromagnetic spectrum information obtained is incomplete, which further aggravates the difficulty of open set target recognition. Therefore, the recognition model not only needs to have the ability to handle unknown targets, but also needs to have scalability and generalization capabilities to deal with data missing.

[0003] In the maritime confrontation scenario, multiple targets operate collaboratively through the communication network, thus forming a complex formation structure and information interaction network. By analyzing the electromagnetic spectrum, the topological structure of the formation can be captured, and then a network model reflecting the communication relationship between the targets can be constructed. In the communication graph model, the targets are regarded as nodes, the links between them are defined as edges, and the electromagnetic information of the targets is used as the attributes of the nodes. In recent years, the Graph Neural Network (GNN) has shown good results in processing graph structure data, and has been applied to multiple fields such as intelligent recommendation, traffic flow prediction, and knowledge graph optimization. To this end, the present invention considers using GNN to learn the deep features of nodes in the formation graph, so as to realize open set target recognition. However, the existing GNN model has limitations in processing complex network structures. For example, the Graph Convolutional Network (GCN) adopts a fixed weighted average method, ignoring the differences in the influence of different neighbor nodes on the central node, thereby affecting the learning and expression ability of the model. GraphSAGE reduces the amount of calculation by sampling neighbor nodes, which may lead to the loss of important information. Especially in graphs with large node degrees, sampling may not capture all important neighbor nodes, thus affecting the accuracy of the model. Graph Attention Network (GAT) improves the information aggregation effect by introducing an attention mechanism to weight the importance of neighboring nodes. However, the attention mechanism may over-focus on certain local structures and ignore other important global information, resulting in a decrease in the generalization ability of the model. Therefore, it is necessary to improve GAT's ability to capture global information, optimize the graph model feature extraction performance, and then improve the perception accuracy of open set targets to achieve a deeper understanding of the battlefield situation. Summary of the invention

[0004] The purpose of the present invention is to provide an electromagnetic spectrum sensing open set target recognition method based on AGGAT.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] An electromagnetic spectrum sensing open set target recognition method based on AGGAT, the specific steps are as follows:

[0007] Step 1: Multimodal feature representation and processing based on graphical models;

[0008] First, the communication features, radar data and movement patterns of the target are extracted based on the electromagnetic spectrum to build a knowledge graph model of the formation. Then, the electromagnetic spectrum data is converted into a feature matrix under the graph model, and the adjacency matrix of the formation is built based on the communication link. Finally, the non-numeric data in the knowledge graph is encoded, and all the numerical data is scaled based on the Min-Max normalization mechanism.

[0009] Step 2: Target feature extraction network construction;

[0010] In order to solve the problem of weak global feature capture ability caused by local structural deviation of GAT, a global attention mechanism is introduced to extract the features of all nodes in the graph; adaptive weights are introduced to dynamically allocate the importance of local and global features to form an adaptive global attention network AGGAT; finally, AGGAT is applied to the formation graph to update the feature information of each target through back propagation;

[0011] Step 3: Design of loss function of target recognition network;

[0012] Firstly, a cross entropy loss function based on the target recognition task is constructed to match the predicted probability distribution of each node with its actual category label. Then, a contrastive loss function suitable for graph attention network is constructed based on the contrastive learning idea to enhance the separability of feature representation by adjusting the similarity between positive and negative samples. Finally, the cross entropy and contrastive loss are combined to form a comprehensive loss function, so as to focus on both the accuracy of node classification and the distinguishing ability of node features during training.

[0013] Step 4: Feature clustering and feature comparison model construction;

[0014] Firstly, a target feature clustering model based on DBSCAN is constructed, which includes defining the neighborhood radius and the minimum number of points, calculating the core points and boundary points according to the density, and dividing the feature nodes into several clusters. Then, the Euclidean distance between the central feature of each cluster and the known target feature is calculated, and the known targets are matched according to the minimum distance. Clusters exceeding the threshold are marked as unknown targets to realize dynamic target recognition.

[0015] Furthermore, the step 1 is specifically as follows:

[0016] Step 1.1: Define the target radar features, communication features, and motion features mined based on the electromagnetic spectrum as the attributes of each target, define the communication links between targets as relationships, and form a target knowledge graph representation model;

[0017] Step 1.2: Convert the target formation represented by the knowledge graph into a graph model, define the target as a node in the graph, define the communication link as an edge, and define the electromagnetic spectrum data as the attribute of the node; suppose there are n targets in the formation, each target has f-dimensional features, then construct the cluster target feature matrix X according to the feature information of the target, the formula is as follows:

[0018]

[0019] At the same time, the adjacency matrix A is constructed according to the communication links between the targets. The formula is as follows:

[0020]

[0021] If target i communicates with target j, then a ij =1, otherwise a ij =0.

[0022] Step 1.3: The target knowledge graph representation model contains non-numerical data of communication protocols and modulation methods, which cannot be directly used for mathematical operations in neural networks; the label encoding mechanism is used to convert them into numerical data; and all numerical data are scaled based on the Min-Max normalization mechanism;

[0023]

[0024] Where x represents the original value of the feature, max(x) and min(x) represent the maximum and minimum values ​​of the feature among all targets, respectively, and x′ represents the normalized target feature value;

[0025] Furthermore, the step 2 is specifically as follows:

[0026] Step 2.1: Based on the GAT model, design an AGGAT model that combines local attention and global attention mechanisms to extract the feature information of each target in the formation graph;

[0027] Step 2.2: AGGAT introduces adaptive parameters to achieve dynamic aggregation of local features and global features; the local attention mechanism is used to capture the relationship between a node and its neighbors; for each node i, the attention weight between it and its neighbor j is calculated; in the local attention mechanism, the feature h of each node i =[x i1 ,x i2 ,...,x if ] is linearly transformed by a learnable weight matrix W; for each pair of neighbors (i, j), the updated features are concatenated and the local attention score e is calculated ij :

[0028] e ij=LeakyReLU(a T [Wh i ||Wh j ])

[0029] Among them, W and a represent the learnable weight vector;

[0030] Local feature representation of node i is the weighted sum of its neighbors’ features:

[0031]

[0032] Among them, α ij is the local attention weight:

[0033]

[0034] Among them, Ne(i) represents the neighbor set of node i;

[0035] The global attention mechanism is used to capture the dependency between two nodes that are far apart. The global attention mechanism generates a global weight for each node, which is used to perform a weighted summation of all nodes. First, for each node i, its global attention weight is calculated

[0036]

[0037] Among them, a global is a learnable vector of global attention;

[0038] The aggregated global features are obtained by weighted summation of all nodes. for:

[0039]

[0040] in, T is the total number of nodes in the graph;

[0041] Step 2.3: Introducing dynamic weight β in AGGAT i , which is used to adaptively adjust the fusion ratio of local information and global information for each node; according to the local information and global graph information of the node, the final representation of the node feature is Defined as:

[0042]

[0043] In order to further enhance the feature expression ability of the model, AGGAT adopts a multi-head attention mechanism; each head independently calculates local and global attention, and then splices the results together; if the network is set to K heads, the target feature based on multi-head attention is finally output for:

[0044]

[0045] Furthermore, the step 3 is specifically as follows:

[0046] Step 3.1: In the object recognition task based on the graph attention network, the cross entropy loss is used to calculate the difference between the model's prediction and the actual label; the mathematical model formula of the cross entropy loss is as follows:

[0047]

[0048] Among them, N represents the number of samples, C represents the number of categories, and y ij represents the jth element in the true label vector of the i-th sample, Represents the jth element in the predicted probability distribution vector of the i-th sample;

[0049] Step 3.2: In the AGGAT-based target recognition model, the InfoNCE loss function is introduced to calculate the feature similarity between different targets and the comparison loss function L Con as follows:

[0050]

[0051] Among them, z i represents the embedding vector of sample i, represents the positive sample embedding vector of sample i, represents the negative sample embedding vector of sample i, τ is the temperature coefficient, sim(z b ,z c ) represents the cosine similarity,

[0052] Step 3.3: Combine the cross entropy loss and InfoNCE loss to form a comprehensive loss function; during the training process, the cross entropy loss is responsible for optimizing the classification performance of the node, while the contrast loss improves the distinguishability of the node features; the final loss function L T The definition is as follows:

[0053] L T =λ 1 L Cro +λ 2 L Con

[0054] Among them, λ 1 and λ 2 Represents the weight coefficient.

[0055] Furthermore, the step 4 is specifically as follows:

[0056] The feature matching based on Euclidean distance quantifies the similarity by measuring the spatial distance between node features. The smaller the distance, the more similar the features are. Suppose the feature vectors of the two targets are u=(u 1 ,u 2 ,...,u m ) and v=(v 1 ,v 2 ,...,v m ), then the characteristic Euclidean distance d(u,v) of two targets is defined as:

[0057]

[0058] Set the threshold of similarity between nodes to ψ; traverse all known target features, match them with the target features to be identified output by the network, and output the minimum Euclidean distance d with the known target type s. s ; if d s <ψ, then the target type is considered to belong to type s, otherwise it is considered to belong to a new category.

[0059] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an electromagnetic spectrum perception open set target recognition method based on AGGAT.

[0060] A computer-readable storage medium stores a computer program / instruction thereon, which, when executed by a processor, implements the steps of an electromagnetic spectrum sensing open set target recognition method based on AGGAT.

[0061] The beneficial effects of the present invention are:

[0062] 1. An open set target recognition model under incomplete data conditions is constructed based on electromagnetic spectrum information. A target graph model is constructed based on target communication, radar and motion features, and a graph attention network and DBSCAN are used to extract and cluster target features. A feature comparison model is introduced to realize the recognition of known types of targets and the clustering of unknown types of targets.

[0063] 2. An improved adaptive graph attention network AGGAT based on local and global attention mechanisms is designed. AGGAT adopts a global attention mechanism so that the final feature representation of each node can take into account both neighborhood information and full graph structure information. In addition, AGGAT introduces adaptive parameters to dynamically balance the weights of local and global information so as to adaptively select information sources under different data distributions, thereby improving the flexibility and representation ability of feature extraction of the graph attention model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is the workflow diagram of open set object recognition;

[0065] Figure 2 It is an open set object recognition network framework;

[0066] Figure 3 It is a knowledge graph representation model of target formation;

[0067] Figure 4 It is the AGGAT network feature update model;

[0068] Figure 5 It is the two-dimensional visualization result of target recognition;

[0069] Figure 6 is the open set object recognition accuracy of different feature extraction models. DETAILED DESCRIPTION

[0070] The present invention is further described below in conjunction with the accompanying drawings.

[0071] according to Figure 1 , the specific implementation of the present invention includes four main steps: multimodal feature representation and processing based on the graph model, target feature extraction network construction, target recognition network loss function design, feature clustering and feature comparison model construction. First, a knowledge graph representing the formation information is constructed based on the multi-dimensional features of the target extracted from the electromagnetic spectrum, and it is converted into a representation form under the graph model through data encoding and normalization; secondly, a global attention mechanism is introduced to construct an adaptive global graph attention network (Adaptive Global GAT, AGGAT) to achieve the comprehensive utilization of local information and global information in the graph, thereby achieving feature extraction of targets in the formation graph, and AGGAT is applied to the feature extraction of targets; subsequently, a target type recognition loss function that comprehensively utilizes supervised learning and contrastive learning is designed, and applied to the training process of target type feature update. Finally, a feature clustering model based on the density-based spatial clustering of applications with noise (DBSCAN) is constructed, and a feature comparison model based on Euclidean distance is constructed to output the type of open set targets. In summary, the network architecture of the open set target recognition model is as follows: Figure 2 shown.

[0072] (1) Multimodal feature representation and processing based on graphical models

[0073] The present invention uses knowledge graph to realize the feature representation of cluster targets. The target radar features, communication features and motion features mined based on the electromagnetic spectrum are defined as the attributes of each target, and the communication links between targets are defined as relationships. The target knowledge graph representation model formed is as follows: Figure 3 As shown. Subsequently, the target formation represented by the knowledge graph is converted into a graph model, the target is defined as a node in the graph, the communication link is defined as an edge, and the electromagnetic spectrum data is defined as the attribute of the node. Assuming that there are n targets in the formation, each target has f-dimensional features, then the cluster target feature matrix shown in formula (1) is constructed based on the feature information of the target.

[0074]

[0075] At the same time, the adjacency matrix shown in formula (2) is constructed based on the communication links between targets. If target i communicates with target j, then a ij =1, otherwise a ij =0.

[0076]

[0077] by Figure 3 For example, the formation includes 2 types of ships (SHIP1 and SHIP2) and 4 types of aircraft (AIR1, AIR2, AIR3 and AIR4), totaling 20 targets. The radar attributes of all targets include carrier frequency, pulse width, pulse amplitude, and pulse repetition interval; the communication attributes of the targets include communication frequency, communication bandwidth, modulation mode, and communication protocol; the motion attributes of the targets include the height and motion speed information of the targets. The target features include different data types. Since non-numerical data such as communication protocol and modulation mode cannot be directly used for mathematical operations in neural networks, the present invention uses a label encoding mechanism to convert them into numerical data. In addition, the data numerical ranges of different features vary greatly, so that some features dominate the model, while the influence of other features will be weakened by one step, resulting in the model being too sensitive to some features, affecting the generalization ability, and increasing the complexity and training difficulty of the model. To this end, the present invention uses the "min-max normalization" mechanism shown in formula (3) to pre-process all features, thereby ensuring that the model training process is more stable and efficient.

[0078]

[0079] Among them, x represents the original value of the feature, max(x) and min(x) represent the maximum and minimum values ​​of the feature among all targets respectively, and x′ represents the normalized target feature value.

[0080] (2) Target feature extraction network construction

[0081] The processed target formation graph model is input into Figure 2 The feature extraction unit shown in the figure realizes embedding update. The present invention designs an AGGAT model that combines local attention and global attention mechanisms to extract feature information of each target in the formation graph. In addition, AGGAT introduces adaptive parameters to realize dynamic aggregation of local features and global features. Its feature update model is as follows: Figure 4 As shown in Figure 2, the local attention mechanism is used to capture the relationship between a node and its neighbors. For each node i, the attention weight between it and its neighbor j is calculated. In the local attention mechanism, the feature h of each node i A linear transformation is performed through a learnable weight matrix W. For each pair of neighbors (i, j), the updated features are concatenated and the local attention score is calculated:

[0082] e ij =LeakyReLU(a T [Wh i ||Wh j ]) (1)

[0083] Among them, W and a represent the learnable weight vectors.

[0084] Then, the attention score e ij Normalized by softmax, the local attention weight α is obtained ij :

[0085]

[0086] Among them, Ne(i) represents the neighbor set of node i.

[0087] Finally, the feature of node i is represented as the weighted sum of its neighbors’ features:

[0088]

[0089] The global attention mechanism is used to capture the global structural information of the entire graph, especially to capture the dependency between two nodes that are far apart. The global attention mechanism generates a global weight for each node, which is used to perform a weighted summation of all nodes. First, for each node i, its global attention weight is calculated:

[0090]

[0091] Among them, a global is a learnable vector of global attention.

[0092] Subsequently, the global attention weights are normalized using softmax:

[0093]

[0094] Where T is the total number of nodes in the graph.

[0095] The aggregated global features are obtained by weighted summation of all nodes:

[0096]

[0097] In addition, AGGAT introduces a dynamic weight β i , which is used to adaptively adjust the fusion ratio of local information and global information for each node. The adaptive weight is obtained by linear transformation of node features:

[0098]

[0099] Among them, W β is a learnable weight parameter.

[0100] According to the local information of the node and the whole graph information, the final representation of the node feature can be defined as:

[0101]

[0102] In order to further enhance the feature expression ability of the model, AGGAT adopts a multi-head attention mechanism. Each head calculates local and global attention independently, and then splices the results together. Assuming that the network is set to K heads, the target feature based on multi-head attention is finally output as:

[0103]

[0104] (3) Loss function design of target recognition network

[0105] In the open set object recognition model based on the AGGAT network, in order to improve the distinguishability of nodes in the embedding space, the present invention constructs a loss function based on the combination of cross entropy loss and contrast loss, which improves the node classification ability while enhancing the feature representation ability of the model.

[0106] Cross entropy loss optimizes the model's prediction ability by measuring the difference between the model's predicted probability distribution and the true label. In the object recognition task based on the graph attention network, cross entropy loss is used to calculate the difference between the model's prediction and the actual label. The mathematical model of cross entropy loss is shown in formula (10).

[0107]

[0108] Among them, N represents the number of samples, C represents the number of categories, and y ij represents the jth element in the true label vector of the i-th sample, Represents the jth element in the predicted probability distribution vector of the i-th sample.

[0109] Contrastive loss is a loss function based on similarity. Its core idea is to improve the distinguishing ability of feature representation by maximizing the distance between similar samples and minimizing the distance between dissimilar samples. In the target recognition model based on AGGAT, the present invention introduces the InfoNCE loss function to calculate the feature similarity between different targets. The contrastive loss function is shown in formula (11).

[0110]

[0111] Among them, z i represents the embedding vector of sample i, represents the positive sample embedding vector of sample i, represents the negative sample embedding vector of sample i, τ is the temperature coefficient, sim(z b ,z c ) represents cosine similarity, and its mathematical model is shown in formula (12).

[0112]

[0113] In order to comprehensively consider the accuracy of node classification and the representation ability of node features, the present invention combines cross entropy loss and InfoNCE loss to form a comprehensive loss function. During the training process, cross entropy loss is responsible for optimizing the classification performance of nodes, while contrast loss improves the distinguishability of node features. The final loss function is shown in formula (13).

[0114] L T =λ 1 L Cro +λ 2 L Con (13)

[0115] Among them, λ 1 and λ 2 Represents the weight coefficient.

[0116] (4) Feature clustering and feature comparison model construction

[0117] The goal of feature clustering is to cluster nodes in the feature space according to similarity, with each cluster corresponding to a potential target category. In open set problems, since there are unknown categories, clustering algorithms can help discover these unknown targets and distinguish them from known categories. Figure 2The feature clustering module shown in the figure adopts the DBSCAN algorithm to implement the classification process based on the target feature. The core idea of ​​DBSCAN is to define clusters based on node density and form clusters by connecting points that can be reached by density. The advantage of the DBSCAN algorithm is that it can find clusters of any shape without specifying the number of clusters in advance. The main parameters of the DBSCAN algorithm are two: the maximum distance Eps of neighbor nodes and the minimum number of neighbors MinPts required to become a core object.

[0118] The DBSCAN-based feature clustering algorithm first marks all feature points as unvisited, and then randomly selects an unvisited point P to mark as visited. Next, the algorithm determines the neighborhood of P, that is, all points whose distance to P is less than or equal to Eps. If there are at least MinPts points in the neighborhood of P, then P is identified as a core object, and a new cluster C is created, and all points in P and its neighborhood are added to C. For each point Q in cluster C, the algorithm repeats the above steps until no new points can be added to C. If P is not a core object, it is marked as a noise point. This process is repeated until all points have been visited.

[0119] The goal of the feature matching module is to compare the clusters obtained by feature clustering with target features of known types, calculate the similarity, and thus identify clusters that match the known categories. Figure 2 The feature matching module shown in the figure uses Euclidean distance to determine the similarity between target types. The feature matching based on Euclidean distance quantifies the similarity by measuring the spatial distance between node features. The smaller the distance, the more similar the features are. Assume that the feature vectors of the two targets are u=(u 1 ,u 2 ,...,u m ) and v=(v 1 ,v 2 ,…,v m ), then the characteristic Euclidean distance of two targets can be defined as:

[0120]

[0121] Set the threshold of similarity between nodes to ψ. Traverse all known target features, match them with the target features to be identified output by the network, and output the minimum Euclidean distance d with the known target type s. s If d min <ψ, then the target type is considered to belong to type s, otherwise it is considered to belong to a new category.

[0122] The effectiveness of the method of the present invention is verified as follows:

[0123] The model proposed in the present invention is divided into two processes: model training and testing in the performance verification stage. In the open set target recognition model training stage, the training process is performed using targets of known types, and the model training adopts the PyTorch framework. In the training stage, the target type is defined as a label, the feature matrix and adjacency matrix after formation conversion are input into the target feature extraction model based on AGGAT, and the embedded features of the target are updated based on the true label of the target. Subsequently, the parameters Eps and MinPts of the clustering algorithm and the feature Euclidean distance threshold ψ of the feature matching model are determined in the verification set. At the same time, the relevant parameters of the feature extraction model, clustering model and matching model are saved. In the model testing stage, an open set target formation with unknown types is input into the trained network, and the types of known targets and clustering results of unknown types are output.

[0124] (1) Feasibility verification

[0125] The present invention uses 1200 target formations to conduct feasibility verification analysis on the proposed model. Each formation contains 18 to 20 targets, and there are 12 target types in total. The characteristics and communication relationships of the targets in each formation are different. Among them, there are 2 types of ships and 4 types of aircraft (SHIP1, SHIP2, AIR1, AIR2, AIR3, AIR4) of known target types, as well as 2 types of ships (SHIP3, SHIP4) and 4 types of aircraft (AIR5, AIR6, AIR7, AIR8) of unknown target types. In the training stage, 100 groups of formations with known target types are used as training samples, and 100 groups of formations with known target types are defined as verification sets to determine the similarity radius of the clustering algorithm and the distance threshold of the feature comparison algorithm. In view of the fact that there is only one target of some types in the formation, the MinPts of the DBSCAN algorithm is defined as 1. The visualization results of the test formation containing 2 types of ships and 4 types of aircraft are shown as follows: Figure 5 As shown in Figure 1, the experimental results show that the proposed model successfully distinguishes between ship and aircraft targets, and correctly divides the targets in each category. For the open set formations of 1000 formations in the test set, the recognition or classification accuracy is shown in Table 1. Analysis of the experimental results shows that the proposed model has a recognition accuracy of 99.79% for known types of targets, and even for unknown types of targets, the classification accuracy exceeds 91%, indicating that the proposed model has certain effectiveness.

[0126] Table 1. Recognition results of open set target test set

[0127]

[0128] (2) Reliability Verification

[0129] Considering that the actual reconnaissance process may be affected by the reconnaissance device, the target information actually obtained is not complete. When the data is missing, the corresponding data is set to 0 after data normalization. The present invention conducts simulation analysis for the case where 10% of the data is missing, and the recognition accuracy of the proposed model is shown in Table 2. The experimental results show that even if the data is missing by 10%, the recognition accuracy of the proposed model for known types of targets is still more than 96%, and the recognition accuracy of unknown targets is also more than 87%, indicating that the proposed model has good reliability.

[0130] Table 2 Open set target test set recognition results (data missing 10%)

[0131]

[0132]

[0133] (3) Method comparison and performance analysis

[0134] To further simulate actual application scenarios, this paper analyzes the performance of the model under different data loss conditions, and compares the proposed AGGAT feature extraction model with MLP and GAT. The selected MLP model is a 2-layer model, and GAT is a two-headed attention model. The loss function uses cross entropy. The comparison results of the recognition accuracy of all target types of each model under different proportions of data loss are shown in Figure 2. Figure 6 As shown. Analysis of the experimental results shows that compared with the MLP feature extraction algorithm that only uses the target's own information, the feature extraction algorithm based on the graph attention network better utilizes the interactive information between targets, thereby effectively improving the accuracy of target recognition. In addition, since the AGGAT model proposed in the present invention comprehensively utilizes the local features and global information of the formation graph structure, and adds a contrast loss function, the nodes can capture more comprehensive information during feature extraction, thereby improving the generalization ability of the model. Therefore, even in the case of 20% missing data, the proposed model comprehensively utilizes feature extraction, feature clustering and feature matching processes, so that the accuracy of open set target recognition still exceeds 84%, indicating that the proposed model has good robustness.

[0135] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An electromagnetic spectrum sensing open set target recognition method based on AGGAT, characterized by: The specific steps are as follows: Step 1: Multimodal feature representation and processing based on graphical models; First, the communication features, radar data and movement patterns of the target are extracted based on the electromagnetic spectrum to build a knowledge graph model of the formation. Then, the electromagnetic spectrum data is converted into a feature matrix under the graph model, and the adjacency matrix of the formation is built based on the communication link. Finally, the non-numeric data in the knowledge graph is encoded, and all the numerical data is scaled based on the Min-Max normalization mechanism. Step 2: Target feature extraction network construction; In order to solve the problem of weak global feature capture ability caused by local structural deviation of GAT, a global attention mechanism is introduced to extract the features of all nodes in the graph; adaptive weights are introduced to dynamically allocate the importance of local and global features to form an adaptive global attention network AGGAT; finally, AGGAT is applied to the formation graph to update the feature information of each target through back propagation; Step 3: Design of loss function of target recognition network; Firstly, a cross entropy loss function based on the target recognition task is constructed to match the predicted probability distribution of each node with its actual category label. Then, a contrastive loss function suitable for graph attention network is constructed based on the contrastive learning idea to enhance the separability of feature representation by adjusting the similarity between positive and negative samples. Finally, the cross entropy and contrastive loss are combined to form a comprehensive loss function, so as to focus on both the accuracy of node classification and the distinguishing ability of node features during training. Step 4: Feature clustering and feature comparison model construction; First, a target feature clustering model based on DBSCAN is constructed, including defining the neighborhood radius and the minimum number of points, calculating the core points and boundary points according to the density, and dividing the feature nodes into several clusters; Subsequently, the Euclidean distance between each cluster center feature and the known target feature is calculated, and the known target is matched according to the minimum distance. Clusters exceeding the threshold are marked as unknown targets to achieve dynamic target recognition.

2. The electromagnetic spectrum sensing open set target recognition method based on AGGAT according to claim 1 is characterized by: The step 1 is specifically as follows: Step 1.1: Define the target radar features, communication features, and motion features mined based on the electromagnetic spectrum as the attributes of each target, define the communication links between targets as relationships, and form a target knowledge graph representation model; Step 1.2: Convert the target formation represented by the knowledge graph into a graph model, define the target as a node in the graph, define the communication link as an edge, and define the electromagnetic spectrum data as the attribute of the node; suppose there are n targets in the formation, each target has f-dimensional features, then construct the cluster target feature matrix X according to the feature information of the target, the formula is as follows: At the same time, the adjacency matrix A is constructed according to the communication links between the targets, and the formula is as follows: If target i communicates with target j, then a ij =1, otherwise a ij =0. Step 1.3: The target knowledge graph representation model contains non-numerical data of communication protocols and modulation methods, which cannot be directly used for mathematical operations in neural networks; the label encoding mechanism is used to convert them into numerical data; and all numerical data are scaled based on the Min-Max normalization mechanism; Among them, x represents the original value of the feature, max(x) and min(x) represent the maximum and minimum values ​​of the feature among all targets respectively, and x′ represents the normalized target feature value.

3. The electromagnetic spectrum sensing open set target recognition method based on AGGAT according to claim 1 is characterized by: The step 2 is specifically as follows: Step 2.1: Based on the GAT model, design an AGGAT model that combines local attention and global attention mechanisms to extract the feature information of each target in the formation graph; Step 2.2: AGGAT introduces adaptive parameters to achieve dynamic aggregation of local features and global features; the local attention mechanism is used to capture the relationship between a node and its neighbors; for each node i, the attention weight between it and its neighbor j is calculated; in the local attention mechanism, the feature h of each node i =[x i1 ,x i2 ,...,x if ] is linearly transformed by a learnable weight matrix W; for each pair of neighbors (i, j), the updated features are concatenated and the local attention score e is calculated ij : e ij =LeakyReLU(a T [Wh i ||Wh j ]) Among them, W and a represent the learnable weight vector; Local feature representation of node i is the weighted sum of its neighbors’ features: Among them, α ij is the local attention weight: Among them, Ne(i) represents the neighbor set of node i; The global attention mechanism is used to capture the dependency between two nodes that are far apart. The global attention mechanism generates a global weight for each node, which is used to perform a weighted summation of all nodes. First, for each node i, its global attention weight is calculated Among them, a global is a learnable vector of global attention; The aggregated global features are obtained by weighted summation of all nodes. for: in, T is the total number of nodes in the graph; Step 2.3: Introducing dynamic weight β in AGGAT i , which is used to adaptively adjust the fusion ratio of local information and global information for each node; according to the local information and global graph information of the node, the final representation of the node feature is Defined as: In order to further enhance the feature expression ability of the model, AGGAT adopts a multi-head attention mechanism; each head independently calculates local and global attention, and then splices the results together; if the network is set to K heads, the target feature based on multi-head attention is finally output for:

4. The electromagnetic spectrum sensing open set target recognition method based on AGGAT according to claim 1, characterized in that: The step 3 is as follows: Step 3.1: In the object recognition task based on the graph attention network, the cross entropy loss is used to calculate the difference between the model's prediction and the actual label; the mathematical model formula of the cross entropy loss is as follows: Among them, N represents the number of samples, C represents the number of categories, and y ij represents the jth element in the true label vector of the i-th sample, Represents the jth element in the predicted probability distribution vector of the i-th sample; Step 3.2: In the AGGAT-based target recognition model, the InfoNCE loss function is introduced to calculate the feature similarity between different targets and the comparison loss function L Con as follows: Among them, z i represents the embedding vector of sample i, represents the positive sample embedding vector of sample i, represents the negative sample embedding vector of sample i, τ is the temperature coefficient, sim(z b ,z c ) represents the cosine similarity, Step 3.3: Combine the cross entropy loss and InfoNCE loss to form a comprehensive loss function; during the training process, the cross entropy loss is responsible for optimizing the classification performance of the node, while the contrast loss improves the distinguishability of the node features; the final loss function L T The definition is as follows: L T =λ1L Cro +λ2L Con Among them, λ1 and λ2 represent weight coefficients.

5. The electromagnetic spectrum sensing open set target recognition method based on AGGAT according to claim 1, characterized in that: The step 4 is specifically as follows: Feature matching based on Euclidean distance quantifies the similarity by measuring the spatial distance between node features. The smaller the distance, the more similar the features are. Assume that the eigenvectors of the two targets are u=(u1,u2,...,u m ) and v=(v1,v2,...,v m ), then the characteristic Euclidean distance d(u,v) of two targets is defined as: Set the threshold of similarity between nodes to ψ; traverse all known target features, match them with the target features to be identified output by the network, and output the minimum Euclidean distance d with the known target type s. s ; if d s <ψ, then the target type is considered to belong to type s, otherwise it is considered to belong to a new category.

6. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images